A Lightweight Underwater Object Detection Algorithm with Adaptive Image Enhancement Based on YOLOv8

Abstract

This paper introduces a lightweight underwater object detection algorithm based on YOLOv8, essential for underwater robotics challenged by environmental complexity and real-time demands. Firstly, to enhance underwater image quality without significantly increasing computational demands, this study introduces an Adaptive Underwater Image Enhancement module utilizing lightweight convolutions and digital filters for dynamic enhancement. Secondly, a Re-parameterized Partial Convolution Block is proposed and integrated, replacing foundational blocks in the baseline model’s architecture, resulting in reduced detection network parameters and enhanced accuracy. Additionally, performance evaluation on the UTDAC dataset demonstrates our model achieving a $46.8 \% \mathrm{mAP}$, marking a 1.6% improvement over the baseline, with a total parameter count of merely 2.81 M. Ablation studies and extended experiments validate the effectiveness and adaptability of the proposed modules. Experimental results show that the model achieves a superior balance between accuracy and processing speed, making it particularly suitable for underwater robotic perception.